Agentic Intent Classifier

Overview

The Agentic Intent Classifier is a multi-head query classification stack for conversational traffic. It is the reference implementation that produces the Agentic Intent Taxonomy decision envelope used throughout the AIP protocol. Given a single user query, the classifier runs three neural classification heads in a single forward pass, resolves IAB content taxonomy via a local embedding index, applies calibration, computes a commercial score, evaluates fallback and policy rules, and emits a schema-validated combined envelope containing model_output, system_decision, and meta. Source: agentic-intent-classifier Hugging Face: admesh/agentic-intent-classifier

Quickstart

from transformers import pipeline

clf = pipeline(
    "admesh-intent",
    model="admesh/agentic-intent-classifier",
    trust_remote_code=True,
)

out = clf("Which laptop should I buy for college?")
print(out["model_output"]["classification"]["intent"])
print(out["model_output"]["classification"]["iab_content"])
print(out["system_decision"]["policy"])

Architecture

The classifier uses a shared DistilBERT encoder with three linear classification heads trained via multitask learning, plus a separate IAB content classifier.

                     query_text
                         │
                   ┌─────┴─────┐
                   │ DistilBERT │  (shared encoder, single forward pass)
                   │  Encoder   │
                   └─────┬─────┘
                         │
              ┌──────────┼──────────┐
              ▼          ▼          ▼
        intent_type  intent_subtype  decision_phase
           head         head            head
              │          │              │
              ▼          ▼              ▼
         calibration  calibration  calibration
              │          │              │
              └──────────┼──────────────┘
                         │
                   ┌─────┴─────┐
                   │    IAB    │  (separate classifier + embedding index)
                   │ Classifier│
                   └─────┬─────┘
                         │
              ┌──────────┼──────────┐
              ▼          ▼          ▼
         commercial   fallback    policy
           score      logic       rules
              │          │          │
              └──────────┼──────────┘
                         │
                         ▼
              schema-validated envelope
        (model_output + system_decision + meta)

Heads

Head	Base Model	Labels	Output
`intent_type`	DistilBERT (shared)	10 labels	High-level intent classification
`intent_subtype`	DistilBERT (shared)	18 labels	Interaction pattern
`decision_phase`	DistilBERT (shared)	7 labels	Commercial funnel stage
`iab_content`	Separate classifier	Full IAB TSV paths	IAB Content Taxonomy 3.1 mapping (tier1–tier4)

The three multitask heads share a single encoder and run in one forward pass. The IAB head runs separately with its own classifier and a local embedding index over taxonomy nodes for retrieval-based reranking.

Taxonomy Labels

`intent.type` (10 labels)

Label	Description
`informational`	Seeking knowledge or understanding
`exploratory`	Open-ended exploration without a specific goal
`commercial`	Evaluating products, services, or solutions
`transactional`	Ready to take a purchase or conversion action
`support`	Seeking help with an existing product or service
`personal_reflection`	Introspective or emotional processing
`creative_generation`	Requesting creative content production
`chit_chat`	Casual conversation with no task goal
`ambiguous`	Intent cannot be determined from context
`prohibited`	Content that violates safety or policy rules

`intent.decision_phase` (7 labels)

Label	Description
`awareness`	First contact with a topic or need
`research`	Actively gathering information
`consideration`	Evaluating options and narrowing choices
`decision`	Ready to choose or commit
`action`	Executing the decision (purchase, signup, etc.)
`post_purchase`	After the transaction — review, setup, feedback
`support`	Ongoing support for an existing relationship

`intent.subtype` (18 labels)

Label	Description
`education`	Learning or understanding a concept
`product_discovery`	Finding new products or services
`comparison`	Comparing alternatives side by side
`evaluation`	Assessing a specific option in depth
`deal_seeking`	Looking for promotions, discounts, or deals
`provider_selection`	Choosing a vendor or provider
`signup`	Creating an account or starting a trial
`purchase`	Buying a product or subscribing
`booking`	Reserving or scheduling a service
`download`	Downloading software, content, or resources
`contact_sales`	Reaching out to a sales team
`task_execution`	Performing a specific operational task
`onboarding_setup`	Setting up or configuring a product
`troubleshooting`	Diagnosing and fixing issues
`account_help`	Account access, settings, or management
`billing_help`	Billing, invoicing, or payment issues
`follow_up`	Continuing a previous conversation thread
`emotional_reflection`	Processing feelings or experiences

`iab_content`

IAB content mapping is derived from every row in the IAB Content Taxonomy 3.1 TSV. The classifier supports tier1 through tier4 with mapping_mode (exact, nearest_equivalent, internal_extension) and mapping_confidence.

Inference Pipeline

The classify_query function is the single entry-point for classification:

Multitask forward pass — Runs the shared DistilBERT encoder once. Decodes logits for all three heads (intent_type, intent_subtype, decision_phase) via calibrated argmax.
Confidence calibration — Each head applies per-label calibration artifacts (when available) to produce calibrated confidence scores with threshold gating.
Commercial score — Computed from intent_type, decision_phase, and subtype using a deterministic heuristic.
IAB content resolution — Runs a separate supervised classifier over full taxonomy paths. Uses a local embedding index for taxonomy-node retrieval and reranking. Outputs tier1–tier4 with mapping_mode and mapping_confidence.
Fallback evaluation — If any head falls below its confidence threshold, fallback metadata is produced with reason (confidence_below_threshold, ambiguous_query, policy_default) and safe defaults.
Policy and opportunity — Applies commercial_score thresholds, sensitivity rules, and regulated-vertical checks to produce monetization_eligibility and opportunity type/strength.
Schema validation — The full envelope is validated against the response schema before returning.

Output Envelope

The classifier produces the Agentic Intent Taxonomy envelope with an additional meta block:

{
  "model_output": {
    "classification": {
      "iab_content": {
        "taxonomy": "IAB Content Taxonomy",
        "taxonomy_version": "3.1",
        "tier1": { "id": "52", "label": "Business and Finance" },
        "tier2": { "id": "53", "label": "Business" },
        "tier3": { "id": "72", "label": "Business I.T." },
        "mapping_mode": "nearest_equivalent",
        "mapping_confidence": 0.93
      },
      "intent": {
        "type": "commercial",
        "subtype": "comparison",
        "decision_phase": "decision",
        "confidence": 0.87,
        "commercial_score": 0.92,
        "summary": "User is evaluating CRM tools for a small team.",
        "component_confidence": {
          "intent_type": {
            "label": "commercial",
            "confidence": 0.91,
            "raw_confidence": 0.88,
            "confidence_threshold": 0.6,
            "calibrated": true,
            "meets_threshold": true
          },
          "intent_subtype": {
            "label": "comparison",
            "confidence": 0.85,
            "raw_confidence": 0.82,
            "confidence_threshold": 0.5,
            "calibrated": true,
            "meets_threshold": true
          },
          "decision_phase": {
            "label": "decision",
            "confidence": 0.89,
            "raw_confidence": 0.86,
            "confidence_threshold": 0.5,
            "calibrated": true,
            "meets_threshold": true
          },
          "overall_strategy": "min_required_component_confidence"
        }
      }
    }
  },
  "system_decision": {
    "policy": {
      "monetization_eligibility": "allowed",
      "eligibility_reason": "commercial_score_above_threshold",
      "decision_basis": "score_threshold",
      "applied_thresholds": {
        "commercial_score_min": 0.7,
        "intent_type_confidence_min": 0.6,
        "intent_subtype_confidence_min": 0.5,
        "decision_phase_confidence_min": 0.5
      },
      "sensitivity": "low",
      "regulated_vertical": false
    },
    "opportunity": {
      "type": "comparison_slot",
      "strength": "high"
    },
    "intent_trajectory": ["decision"]
  },
  "meta": {
    "system_version": "0.9.0",
    "calibration_enabled": true,
    "iab_mapping_is_placeholder": false
  }
}

`meta` block

Field	Type	Description
`system_version`	string	Classifier version
`calibration_enabled`	boolean	Whether any head used calibrated confidence
`iab_mapping_is_placeholder`	boolean	Whether IAB mapping used placeholder fallback

`component_confidence`

Per-head confidence details are included inside model_output.classification.intent.component_confidence. Each head reports:

Field	Type	Description
`label`	string	Predicted label
`confidence`	number	Calibrated confidence (`0.0`–`1.0`)
`raw_confidence`	number	Pre-calibration confidence
`confidence_threshold`	number	Minimum threshold for this head
`calibrated`	boolean	Whether calibration artifacts were applied
`meets_threshold`	boolean	Whether confidence meets the threshold

API

The classifier exposes a local HTTP API via demo_api.py:

`POST /classify`

Request:

{ "text": "best CRM for small teams" }

Response: Full classification envelope (see Output Envelope above).

`GET /health`

Returns head readiness and calibration status:

{
  "status": "ok",
  "system_version": "0.9.0",
  "heads": [
    { "head": "intent_type", "model_path": "...", "calibration_path": "...", "ready": true, "calibrated": true },
    { "head": "intent_subtype", "model_path": "...", "calibration_path": "...", "ready": true, "calibrated": true },
    { "head": "decision_phase", "model_path": "...", "calibration_path": "...", "ready": true, "calibrated": true },
    { "head": "iab_content", "model_path": "...", "calibration_path": "...", "ready": true, "calibrated": true }
  ]
}

`GET /version`

{
  "system_version": "0.9.0",
  "api_schema_version": "2026-03-22"
}

What The Classifier Does

Runs three classifier heads (intent_type, intent_subtype, decision_phase) via a shared multitask DistilBERT encoder
Resolves iab_content through a local embedding index over taxonomy nodes plus label/path reranking
Applies calibration artifacts when present
Computes commercial_score from intent type, decision phase, and subtype
Applies fallback when confidence is too weak or policy-safe blocking is required
Emits a schema-validated combined envelope conforming to the Agentic Intent Taxonomy

What The Classifier Does Not Do

It is not a multi-turn memory system
It is not a production-optimized low-latency serving path (use the operator’s Groq-based classifier for production latency targets)
It is not yet trained on large real-traffic human-labeled intent data
Combined decision logic is heuristic, though materially stronger than rule-based baselines

Relationship to PlatformRequest

In the AIP protocol, the classifier can be used in two ways depending on the classification input mode:

Mode	Who runs the classifier	How it connects
Operator-derived	The operator runs the classifier on `interaction.input.query_text`	Operator uses the classifier output to populate `policy` and drive auction decisions
Platform-derived	The platform runs the classifier locally and sends `provided_signals`	Operator validates the signals via `signal_validation` and may normalize scores

The classifier output envelope (model_output + system_decision) maps directly to the structures in the platform request policy block and the signals.intent block.

Agentic Intent Taxonomy — The decision envelope schema this classifier produces
Platform Request — How classification feeds into the platform request
Hugging Face Model — Hosted model for quick testing

Welcome

Get Started

AI Platforms

Publisher Access Governance (PAG)

Common

Overview

Quickstart

Architecture

Heads

Taxonomy Labels

`intent.type` (10 labels)

`intent.decision_phase` (7 labels)

`intent.subtype` (18 labels)

`iab_content`

Inference Pipeline

Output Envelope

`meta` block

`component_confidence`

API

`POST /classify`

`GET /health`

`GET /version`

What The Classifier Does

What The Classifier Does Not Do

Relationship to PlatformRequest

Welcome

Get Started

AI Platforms

Publisher Access Governance (PAG)

Common

​Overview

​Quickstart

​Architecture

​Heads

​Taxonomy Labels

​intent.type (10 labels)

​intent.decision_phase (7 labels)

​intent.subtype (18 labels)

​iab_content

​Inference Pipeline

​Output Envelope

​meta block

​component_confidence

​API

​POST /classify

​GET /health

​GET /version

​What The Classifier Does

​What The Classifier Does Not Do

​Relationship to PlatformRequest

​Related

Overview

Quickstart

Architecture

Heads

Taxonomy Labels

`intent.type` (10 labels)

`intent.decision_phase` (7 labels)

`intent.subtype` (18 labels)

`iab_content`

Inference Pipeline

Output Envelope

`meta` block

`component_confidence`

API

`POST /classify`

`GET /health`

`GET /version`

What The Classifier Does

What The Classifier Does Not Do

Relationship to PlatformRequest

Related